Embed Size (px)
Citation preview
Tracking the Evolution and Diversity in Network Usage ofSmartphones
Kensuke FukudaNational Institute of Informatics
Hirochika AsaiUniversity of Tokyo
Kenichi NagamiIntec
ABSTRACTWe analyze the evolution of smartphone usage from a datasetobtained from three, 15-day-long, user-side, measurementswith over 1500 recruited smartphone users in the GreaterTokyo area from 2013 to 2015. This dataset shows usersacross a diverse range of networks; cellular access (3G toLTE), WiFi access (2.4 to 5GHz), deployment of more pub-lic WiFi access points (APs), as they use diverse applica-tions such as video, file synchronization, and major softwareupdates.
Our analysis shows that smartphone users select appropri-ate network interfaces taking into account the deployment ofemerging technologies, their bandwidth demand, and theireconomic constraints. Thus, users show diversity in bothhow much traffic they send, as well as on what networksthey send it. We show that users are gradually but steadilyadopting WiFi at home, in offices, and public spaces overthese three years. The majority of light users have beenshifting their traffic to WiFi. Heavy hitters acquire morebandwidth via WiFi, especially at home. The percentage ofusers explicitly turning off their WiFi interface during theday decreases from 50% to 40%. Our results highlight thatthe offloading environment has been improved during thethree years, with more than 40% of WiFi users connectingto multiple WiFi APs in one day. WiFi offload at officesis still limited in our dataset due to a few accessible APs,but WiFi APs in public spaces have been an alternative tocellular access for users who request not only simple con-nectivity but also bandwidth-consuming applications suchas video streaming and software updates.
Categories and Subject DescriptorsC.2.3 [Computer-Communication Networks]: NetworkOperations—Network monitoring
General TermsMeasurement
Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than theauthor(s) must be honored. Abstracting with credit is permitted. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specific permissionand/or a fee. Request permissions from [email protected]’15, October 28–30, 2015, Tokyo, Japan.Copyright is held by the owner/author(s). Publication rights licensed to ACM.ACM 978-1-4503-3848-6/15/10 ...$15.00.DOI: http://dx.doi.org/10.1145/2815675.2815697.
0
500
1000
1500
2000
2500
3000
3500
4000
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
RBB user download
Cellular user download(3G+LTE)
Tra
ffic
volu
me [G
bps]
Year
Figure 1: Growth in residential broadband and cellular traf-fic in Japan [34].
KeywordsSmartphone traffic; LTE; WiFi; WiFi offload
1. INTRODUCTIONIn recent years the deployment of high-speed Internet at
home and bandwidth-consuming applications have been driv-ing the growth of residential broadband traffic. Followingthis growth, smartphone traffic has had a large impact oncellular providers. Figure 1 illustrates the growth in the to-tal residential broadband and cellular (3G+LTE) downloadtraffic volumes in Japan [34]. The total residential broad-band traffic was estimated from the traffic volume measuredat the customer edges of six Japanese commercial Internetservice providers (ISPs). The total cellular traffic was mea-sured in the backbones of four Japanese cellular providerscovering most cellular users in Japan. The cellular trafficvolume from smartphones accounted for 20% of the resi-dential broadband traffic volume at the end of 2014. Onereason for this trend is the rapid deployment of smartphoneswith rich applications. Most Japanese smartphone users arecharged a flat rate with a soft bandwidth cap where user’smaximum bandwidth is limited (e.g., 128kbps) during peakhours for a few days if the previous three days downloadvolume surpasses a threshold (e.g., 1GB); thus, most lightusers are not concerned about saving their cellular trafficvolume.
253
A crucial issue for cellular providers is the limited resourceof channels in cellular networks. Cellular providers employseveral methods to minimize quality of service/quality ofexperience (QoS/QoE) degradation. One commonly usedtechnique to mitigate this issue is to apply a bandwidth capto heavy hitters. Second, providers are deploying new capac-ity such as the new 4G LTE network reaching full coverage inJapanese cities. Finally, they are currently encouraging theircustomers to offload cellular traffic to WiFi networks (WiFioffloading). To enable WiFi offloading, cellular providers aredeploying new free WiFi access points (APs) for their cus-tomers in public places (e.g., cafes, metro stations, streets,and airports). Furthermore, they have begun to give cus-tomers free home WiFi routers with the expectation thatthese will increase offloading at home to residential broad-band. The Japanese government also plans to promote thedeployment of more free APs in public spaces towards the2020 Olympic Games in Tokyo [35]. This promotion is ex-pected with the goal of supporting Internet access for foreignvisitors. Rapid deployment of better WiFi coverage enablescellular providers to offload 3G/LTE traffic to WiFi net-works in many places if they appropriately manage to leadtheir users to these networks.
The success of these WiFi deployment hinges on this keyquestion: how do smartphone users select a network fromthe alternatives available to them? Prior work has not sup-ported a straightforward answer to this question due to thedifficulty in tracking all traffic flows upstream.
To answer this question, we characterize smartphone traf-fic behavior from measurements at the user-side. We previ-ously gave preliminary analysis of a two-days-long Androidsmartphone measurement in [18]. Here we go further, pre-senting analysis of three measurements, each 15 days long,covering over 1500 Android and iPhone users in the GreaterTokyo area from 2013 to 2015. We characterize the evolu-tion of smartphone usage in terms of cellular access, WiFiaccess, geolocation, application, device OS, and bandwidthcap.
The main findings of the paper are as follows. (1) Smart-phone traffic is becoming the main player in home networks(§ 3.1, § 3.2, and § 4). We estimate that the daily offloadedtraffic volume for light users accounts for 12% of residentialbroadband light user traffic volume. Also, the estimatedtotal offloaded traffic volume accounts for 28% of the totalbroadband traffic volume. (2) WiFi traffic volume has in-creased more than cellular traffic volume due to the largecontribution of heavy hitters. Even for light users (i.e., me-dian), the WiFi traffic is higher than that of cellular trafficas of 2015 (§ 3.2, and § 4). (3) Users properly select net-work interfaces; the number of users explicitly turning offWiFi has decreased by 10% in three years (§ 3.3). Also, thenumber of associated WiFi networks per user has increased.As of 2015, 40% of WiFi users are associated with at leasttwo APs in one day (§ 3.4). The traffic volume to publicand office WiFi APs, however, only accounts for 2% of thetotal WiFi traffic volume. (4) The signal strength of theassociated WiFi network is generally high; however, 12% ofthe public WiFi networks exhibit subpar quality (§ 3.4, and§ 3.5). We further confirm the rapid deployment of 5GHzAPs in public WiFi networks though dominant APs at homeand office are still 2.4GHz APs. (5) Bandwidth-consumingapplications (e.g., video, online storage) have become morepopular (§ 3.6) in WiFi networks. We see indications that
year duration #And #iOS #total %LTE2013 07 Mar - 22 Mar 948 807 1755 25%2014 28 Feb - 22 Mar 887 789 1676 70%2015 25 Feb - 25 Mar 835 781 1616 80%
Table 1: Overview of datasets.
users use such bandwidth-consuming applications on publicWiFi networks (§ 3.7). Also, software updates require WiFinetworks by default, thus the timing of software update canbe delayed for users without home APs. For security-criticalupdates, this delay may leave users vulnerable longer.
Overall, we characterize the diverse network usage of smart-phones under certain demands (e.g., rich applications) andconstraints (e.g., bandwidth cap and network availability).Our results show slow but clear growth of WiFi traffic of-floading during these three years. In particular, the deploy-ment of public WiFi networks provides users both simplenetwork connectivity and also more bandwidth for bandwidth-intensive applications such as video streaming and softwareupdates. However, we find that WiFi offloading at offices isstill limited.
2. METHODOLOGY AND DATASETWe developed measurement software for Android1 and
iPhone2. The software supports iPhone 4-6 series for iOS,and Android devices with OS version > 2.1. It runs in thebackground and records several pieces of device informationsuch as byte and packet counts per network interface, appli-cation information, battery status, geolocation information,network information (e.g, BSSID, ESSID for WiFi), and aunique random device ID. While most information is similaracross both OSes, there are some differences due to what in-formation is available. Android OS reports application cate-gories and their total traffic volume, but iOS has no interfaceto obtain the traffic volume per application. Similarly, An-droid OS reports non-associated WiFi AP information aswell as associated WiFi APs if the WiFi interface is turnedon, whereas iOS only reports the associated AP information.In concern for energy use and privacy, we report only coarsegeolocation (5km precision) and do not report throughput.The software collects statistics every 10 minutes and uploadsthis data to a central server. If the upload fails the softwarecaches the data and sends it later.
We conducted three measurement campaigns in March of2013, 2014, and 2015 (Table 1)3. Each campaign used anindependently recruited group of subjects from the greaterTokyo area, with more than 800 Android and 700 iPhoneusers. Each recruiting and selection process was done bya marketing research company in consideration of the mar-ket share of major Japanese cellular providers. During thecampaigns, a wide variety of users were selected as listed inTable 2, and they were requested to use their smartphones asusual. The male-female ratio of the users was about 50:50.We also conducted a user survey at the end of the measure-ments in order to better understand user’s behavior that
1https://play.google.com/store/apps/details?id=com.inetcore.linkspeed2https://itunes.apple.com/jp/app/li-yong-shi-tai-diao-zha/id597320740?l=en&mt=83Our work has been institutionally reviewed and approvedas human-subjects research (NII260128, NII270129).
254
Occupation Percentage2013 2014 2015
government worker 2.1 3.4 2.4office worker 20.0 20.1 23.6
engineer 16.7 14.7 16.6worker (other) 12.8 13.7 13.2
professional 2.4 2.0 2.8self-owned business 6.1 6.7 5.6
part timer 9.0 10.1 10.6housewife 15.0 14.2 13.3
student 9.6 8.3 2.7other 6.3 6.8 7.1
Table 2: User survey: user demographics.
cannot be seen in the network data (see details in § 4.2).The number of analyzed unique device IDs was about 1600-1700 over three years. This number includes non-recruitedusers who installed the measurement software from respec-tive app stores.
We intend to characterize users behaviors depending ontheir exchanged traffic volume. Throughout the paper, werefer to light users as those whose daily download trafficranges from the 40th to 60th percentiles, and heavy hittersas users whose daily download traffic is ranked in the top5%. Note that as daily user traffic volume is highly variable,one user may be a light user one day and heavy hitter onanother.
To focus on typical use, we cleaned our data to avoiddistortion due to two atypical events. First, we removedtethering traffic data from the datasets since such data havedifferent traffic characteristics.
Second, we removed traffic relating to the iOS update in2015 from our main analysis. Apple updated iOS(to iOS8.2) during our measurement in 2015. We identified eachdevices update (around March 10th, 2015) and ignored alluser traffic data in the day and the next day of the updatefor our main analysis due to its huge data size (over 500MB).We discuss traffic from this update specifically in § 3.7.
Possible measurement biases: An online market re-search company selected these participants from pooled users,so they are likely more advanced users than the averagesmartphone user. The availability of public WiFi networksin metro stations, cafes, shops, and on the streets is higherin Tokyo than in different areas. Most commuters in thisarea use public transportation (e.g., trains, subways, andbuses) rather than personal cars. Thus, the probability toencounter public WiFi networks is likely high, and resultingWiFi traffic volume in public spaces is also high.
3. ANALYSIS OF SMARTPHONE TRAFFIC
3.1 Aggregated traffic behaviorWe first examine the aggregated traffic behavior of cellular
(3G and LTE) and WiFi networks in our datasets. Figure 2indicates the weekly variations in the aggregated traffic vol-ume in March 2015. The notations TX and RX in the figureare traffic volumes from and to smartphones, respectively.As expected, the WiFi volume exceeds the cellular volume.Thus the traditional measurements of cellular traffic such asFigure 1 underrepresent all smartphone activity due to WiFioffload. The ratio of WiFi traffic to the total traffic increasesover time from 59% in 2013 to 67% in 2015. Also, the cellu-
0
5
10
15
20
25
30
35
40
45
50
28 Sat 01 Sun 02 Mon 03 Tue 04 Wed 05 Thu 06 Fri 07 Sat
Tra
ffic
volu
me [M
bps]
Date (Mar. 2015, JST)
Cellular TXCellular RX
WiFi TXWiFi RX
Figure 2: Aggregated traffic volume.
lar traffic is mainly composed of LTE traffic accounting for32% in 2013 but 80% in 2015 (see also Table 1).
We see that network selection depends strongly on thetime of day, with cellular traffic peaks corresponding to com-mute times and WiFi peaks to evening times at home. Threetraffic peaks in cellular RX are affected by human activi-ties in the morning (8am), noon (12am), and evening (7-9pm). Morning and evening peaks corresponded to the peaktime of commuting mainly by public transportation in theGreater Tokyo area. In contrast, major peaks of the WiFiRX (11pm-1am) are right after the evening peak of cellularRX (9pm), though we also confirm peaks for morning andnoon in WiFi RX. Cellular traffic on weekends is smallerthan that on weekdays, while WiFi traffic is the opposite.Thus, these observations clearly present different temporalusage of WiFi and cellular networks.
3.2 Daily user traffic volumeWe next focus on the daily traffic usage pattern per user.
Our goal is to understand the impact of light users and heavyhitters by characterizing traffic distribution for each user asit changes over three years.
Daily total traffic: We first show that daily trafficincreases each year, and users show a large variation in howmuch they send. Figure 3 shows the CDF of the daily trafficvolume starting from midnight per user per day in a semi-logplot. Note that a single user appears one time for each dayin the trace. We omitted users that downloaded less than0.1MB. The shape of the curves is close to a unimodaldistribution. The traffic volumes obviously increase overtime, and the RX traffic volumes are about five times largerthan the TX traffic volumes. While these two observationshave been reported in residential broadband traffic [10], theresults clearly show the client-server type user behavior.
Daily cellular and WiFi traffic: Next, we present thatdaily WiFi and cellular traffic volumes per user are largelydistributed. Figure 4 presents the CDFs of the daily usertraffic volume for cellular and WiFi network interfaces in2015. As expected, the download and upload traffic volumesare highly skewed. On one hand, 8% of cellular interfacesand 20% of WiFi interfaces do not send and receive anydata. On the other hand, the data shows that nearly allusers respect traffic caps, with only 1.4% of users exceeding
255
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 10 100 1000
CD
F
Daily traffic volume [MB]
All RX13,14,15
All TX13,14,15
Figure 3: CDFs of daily total trafficvolume per user.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.1 1 10 100 1000
CD
F
Daily traffic volume [MB]
2015WiFi RX
WiFi TX
Cellular RX
Cellular TX
Figure 4: CDFs of daily traffic volumeper type.
2015
10-2
10-1
100
101
102
103
Cellular [MB]
10-2
10-1
100
101
102
103
WiF
i [M
B]
0
10
20
30
40
50
60
2015
Figure 5: Daily traffic volume per user.
median 2013 2014 2015 AGRAll 57.9 90.3 126.5 48%
Cell 19.5 27.6 35.6 35%WiFi 9.2 24.3 50.7 134%mean 2013 2014 2015 AGR
All 102.9 179.9 239.5 53%Cell 42.2 58.5 71.5 30%
WiFi 60.7 121.5 168.1 66%
Table 3: Daily download traffic volume per user (MB/day)and annual growth rate.
the 1GB per 3-day soft bandwidth cap. We examine thedetails of the bandwidth cap in § 3.8. When unconstrainedby limits, we see a much longer tail of heavy users—the topheavy hitter downloaded 11GB in one day.
Traffic growth: We emphasize the difference of growthin median and mean daily user traffic volume; they reflecttraffic behavior of light users and heavy hitters, respectively.To investigate traffic growth, we list the median and meandaily user download traffic volumes in Table 3. The mediantraffic indicates the typical light user’s traffic volume, andthe mean traffic reflects traffic from all users, but is biased byheavy hitters. The annual growth rates (AGRs) are obtainedby linear fit.
We see that users download more traffic via WiFi in lateyears. In fact, the traffic growth of median WiFi RX issignificant while the median cellular RX is higher than themedian WiFi RX in 2013. Thus, light users overlook the im-portance of WiFi offloading in 2013, but they change theirbehavior depending on the increase in traffic volume. Sim-ilarly, the growth of the mean WiFi traffic volume is high.Heavy hitters are also heavier in WiFi than cellular in meantraffic volume.
3.3 WiFi vs. CellularTo assess the impact of traffic offloading we next exam-
ine of how users select between WiFi and cellular networkinterfaces.
3.3.1 Aggregated viewWiFi-intensive and cellular-intensive users: We in-
troduce three types of users (cellular-intensive, WiFi-intensive,and mixed user) from the usage of network interfaces. Fig-ure 5 shows a heat map of traffic volumes (log-log scale), for
0
0.2
0.4
0.6
0.8
1
Sat Sun Mon Tue Wed Thu Fri Sat
WiF
i-tr
affic
ratio
(a) WiFi-traffic ratio 20132015
0
0.2
0.4
0.6
0.8
1
Sat Sun Mon Tue Wed Thu Fri Sat
WiF
i-user
ratio
(b) WiFi-user ratio 20132015
Figure 6: (a) WiFi-traffic ratio and (b) WiFi-user ratio.
each user and each day by network (cellular and WiFi). Thecolor bar indicates the number of users per day for a specifictraffic ratio. The diagonal line indicates reference users whodownloaded equal amounts of data from both cellular andWiFi networks.
We find the existence of three typical user types; Cellular-intensive users do not use their WiFi interfaces due to noavailable WiFi APs or no WiFi configuration. WiFi-intensiveusers only rely on their WiFi interfaces to avoid their cel-lular traffic fees. Mixed users select both cellular and WiFinetwork interfaces. The figure indicates that 22% of usersare cellular-intensive and 8% are WiFi-intensive users. Wesee a decrease in cellular-intensive users: from 35% in 2013to 22% in 2015, while the fraction of WiFi-intensive usersis stable during the three years. Interestingly, both WiFi-intensive and cellular-intensive users show a wide spread intraffic volume; thus either can be heavy hitters.
User-level WiFi offloading: This data shows that mixed-network users often offload traffic to WiFi networks. Usersabove the diagonal are evidence of offloading, since theirWiFi traffic exceeds their cellular traffic. The concentrationof users above the diagonal shows WiFi offloading is com-mon. Although 55% of mixed users are above the diagonal,a large number of users remain below it, suggesting WiFiusage can be further improved, especially for high cellulartraffic volume users.
256
0
0.2
0.4
0.6
0.8
1
Sat Sun Mon Tue Wed Thu Fri Sat
WiF
i-tr
affic
ratio
(a) WiFi-traffic ratio (2013) heavylight
0
0.2
0.4
0.6
0.8
1
Sat Sun Mon Tue Wed Thu Fri Sat
WiF
i-tr
affic
ratio
(b) WiFi-traffic ratio (2015) heavylight
Figure 7: WiFi-traffic ratio (2013 and 2015).
3.3.2 WiFi-traffic ratio and WiFi-user ratioWe introduce two metrics to highlight the benefit of us-
ing WiFi networks: The WiFi-traffic ratio is defined as theWiFi download traffic volume divided by the total downloadtraffic volume in one-hour time bins. A ratio close to 1.0 in-dicates that most traffic volume is downloaded via WiFi.The second metric is the WiFi-user ratio, defined as the ra-tio of the number of users associating to WiFi networks ineach time bin. A ratio close to 1.0 means that, at that time,most users are using WiFi.
We see that the WiFi-traffic ratio increases showing in-creasing use of WiFi offloading. Figure 6(a) compares theWiFi-traffic ratio over each day of the week in 2013 with2015. We observe that the WiFi-traffic ratio varies between0.4 and 0.9 with clear diurnal trend. WiFi traffic is the high-est from 11pm to 2am and the lowest in weekday afternoons.This temporal pattern differs from the simple WiFi trafficvolume (Figure 2). Traffic penetration via WiFi is commonat night though cellular traffic is still dominant in some timeperiods during the daytime. This data also confirms thatWiFi offload is increasing, with the mean WiFi-traffic ratiogrowing to 0.71 in 2015 from 0.58 in 2013.
WiFi-user ratio shows that more users connect to WiFinetworks in later years, however only 50% of users connect toWiFi during peak hours. We examine the temporal variationin the WiFi-user ratio in Figure 6(b). The number of userspeaks from 9pm to 2am, while 10am to 6pm is the off-peak.The mean WiFi-user ratio also increases from 32% in 2013to 48% in 2015. Thus, not only heavy hitters but also lightusers recognize the benefit of connecting to WiFi. However,only 50% of devices use WiFi even in the peak time.
3.3.3 Difference in usersWe now highlight the difference in the WiFi-traffic ratio
and WiFi-user ratio between light users and heavy hitters.We emphasize that in 2015, heavy hitters offload most oftheir traffic volume to WiFi, and that the ratio of offloadtraffic for light users also increases over the years. Figure 7represents changes in WiFi-traffic ratio from 2013 to 2015.Heavy hitters already offload most of their traffic to WiFiand this trend shows no obvious diurnal pattern in 2013(mean: 73%) and 2015 (mean: 89%). For light users, theratio increased over the three years (from 42% to 52% inmean), particularly its diurnal pattern is much clearer in
0
0.2
0.4
0.6
0.8
1
Sat Sun Mon Tue Wed Thu Fri Sat
WiF
i-user
ratio
(a) WiFi-user ratio (2013) heavylight
0
0.2
0.4
0.6
0.8
1
Sat Sun Mon Tue Wed Thu Fri Sat
WiF
i-user
ratio
(b) WiFi-user ratio (2015) heavylight
Figure 8: WiFi-user ratio (2013 and 2015).
0 0.1 0.2 0.3 0.4 0.5 0.6
Sat Sun Mon Tue Wed Thu Fri Sat
Ratio o
f users
(a) Android (2013)
WiFi-userWiFi-off
WiFi-available
0 0.1 0.2 0.3 0.4 0.5 0.6
Sat Sun Mon Tue Wed Thu Fri Sat
Ratio o
f users
(b) Android (2015)
WiFi-userWiFi-off
WiFi-available
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Sat Sun Mon Tue Wed Thu Fri Sat
Ratio o
f users
(c) IOS (2013&2015)WiFi-user (2013)WiFi-user (2015)
Figure 9: Ratio of users: (a) Android 2013, (b) Android2015, and (c) iOS 2013 & 2015.
2015, characterized by stable activity on weekends and typ-ical daily trend on weekdays.
We see a large increase of WiFi-user ratio especially forheavy-hitters during the three years. Figure 8 illustratesthe increase of the WiFi-user ratio. The difference betweenheavy hitters and light users is dispersed over the threeyears. In 2013, about 60% of heavy hitters connect to WiFiduring their peak traffic time (mean: 51%). In 2015, morethan 80% of heavy hitters use WiFi during their peak time(mean: 68%). These results suggest that some heavy hittersin 2013 are mainly connected to cellular networks, but thistrend weakens as traffic volume increases over time.
3.3.4 Difference in device OSWe show that there are differences in offloading patterns
correlated with device OS. We first define more detailed cel-lular and WiFi usage to quantify the potential opportunityof WiFi offloading. For Android devices, we categorize usersinto three groups, as shown in Figure 9(a). The first categoryis users connecting their device to WiFi networks (WiFi-user ; pink curve). The second category consists of WiFi-offusers whose WiFi interfaces are off (blue curve). The thirdconsists of WiFi-available users who use cellular interfacesas the primary link but still turn on their WiFi interface(green curve). Note that the WiFi-off users explicitly turn
257
off the WiFi interface and the WiFi-available users do notconnect to any WiFi AP.
We confirm that 40% of Android devices explicitly turnoff their WiFi interface, but this ratio decreases during thethree years. The ratio of the WiFi-available users is sta-ble at around 0.25 in both 2013 and 2015, meaning thata quarter of the users could offload their traffic to WiFi ifany appropriate APs are discovered. We discuss these userslater in § 3.5. In comparison, the variation in the ratio ofWiFi-off users is opposite of that of WiFi users; the peaktime is business hours (10am-6pm). Our results show thatnearly 50% of Android users explicitly turn off their WiFidevices during the day in 2013 (blue curve). However, thesituation of the low WiFi usage improves in 2015, demon-strating that the ratio of WiFi-off users drops from 50% to40% (Figure 9(b)).
WiFi connectivity of iOS is higher than that of Android.For iOS devices, the software does not report detailed in-formation about the WiFi interface. However, we concludethat iOS devices connect to WiFi 30% more than do Androiddevices, as shown in Figure 9(c). We also confirm that thereis no difference in the WiFi-user ratios among three cellu-lar carriers providing iPhones. Thus, WiFi-user ratio differsbetween the two device OSes rather than cellular carriers.
3.4 Usage of WiFi networksNext we illustrate the results of user adoption of available
WiFi networks at home, in offices, and in public places.
3.4.1 Home, public, and other WiFi networksWe first define the locations of WiFi APs users connected
to. We identify each WiFi AP users associate with by its(BSSID, ESSID) pair (the MAC address of AP and its net-work name). We then categorize WiFi networks into thesethree types:
Home: We identify home locations as the most common(BSSID, ESSID) pair to which each device connects duringat least 70% of the time between 10pm and 6am in oneday. Note that users do not always have at least one homenetwork. In fact, the percentage of users with estimatedhome AP increases over time; 66% in 2013, 73% in 2014 and79% in 2015. These numbers are roughly consistent withthe result of our user survey in Table 8. Table 4 lists thenumbers of detected APs.
Public: We identify public networks based on well knownESSID names (e.g., 0000docomo, 0001softbank, eduroam).These services are often deployed by Japanese cellular providersfor their customers (§ 1), or by free/commercial WiFi providers(e.g., 7Spot and Metro Free Wi-Fi). We categorize FONAPs that use a public ESSID at home as a home networkbecause we find that some users connected to the public ES-SID over 24 hours instead of using the default private ES-SID by FON. The number of detected public APs doublesin three years as shown in the table.
Other: Other associated pairs are mainly located at of-fices or mobile WiFi APs. It also includes some open APsprovided by shops and hotels. We further estimate APs atoffices when they (1) mainly connect between 11am and 5pmon weekdays and (2) are not otherwise classified as home,public or mobile APs. The number of estimated office APsis stable over the years.
This classification shows a wide deployment of APs espe-cially in public spaces. Figure 10 visualizes the locations
type 2013 2014 2015home 1139 1223 1289public 5041 9302 10481other 545 673 664(office) 166 168 166total 6725 11198 12434
Table 4: Number of estimated APs.
0
50
100
150
200
250
300
Yokohama
Chiba
Narita
Tokyo
Saitama
Kawasaki
HachiojiFunabashi
Odawara Yokosuka
(a) home 2013
0
50
100
150
200
250
300
Yokohama
Chiba
Narita
Tokyo
Saitama
Kawasaki
HachiojiFunabashi
Odawara Yokosuka
(b) public 2013
0
50
100
150
200
250
300
Yokohama
Chiba
Narita
Tokyo
Saitama
Kawasaki
HachiojiFunabashi
Odawara Yokosuka
(c) home 2015
0
50
100
150
200
250
300
Yokohama
Chiba
Narita
Tokyo
Saitama
Kawasaki
HachiojiFunabashi
Odawara Yokosuka
(d) public 2015
Figure 10: Number of associated unique APs per 5km cell:(a) home in 2013, (b) public in 2013, (c) home in 2015, and(d) public in 2015.
and densities of associated APs in the Greater Tokyo area.Each cell is a 5km square, with color showing identified APdensity. The four quadrants of the figure show home andpublic distribution in 2013 and 2015. Home networks arewidely dispersed, reflecting many residential areas. Thereare only minor differences between the two datasets.
This data suggests that coverage of public WiFi networksis broad in the Greater Tokyo. The maps of public networkshighlight strong densities in the downtown Tokyo area. Thenumber of associated public networks is over 300 in the high-est density areas (Shinjuku and Shibuya areas). We alsoconfirm the wide coverage of public APs even far from thedowntown area. By comparing the two datasets, further-more, high-density areas are more dispersed in 2015. Forexample, the number of cells with at least one AP increasesfrom 229 to 265, and that of cells with larger than 100 APsincreases from 10 to 23. This result suggests that users canpotentially connect to widely deployed public WiFi APs ifdevices are appropriately configured.
However, the contribution of public and office APs is stillsmall when measured by traffic volume. Figure 11 representsthe weekly traffic variations for home, public, and office in2013 and 2015. The major contribution of WiFi traffic vol-ume is home networks (95% of the total volume). The publicand office WiFi traffic volumes are much smaller (4% of thetotal volume), though those volumes double during this pe-riod. The diurnal patterns of public and office WiFi arein opposition to that of home WiFi. These clear patternsin 2015 emphasize the increase of users associated to theseWiFi networks.
258
0
2
4
6
8
10
12
14
16
18
09 Sat 10 Sun 11 Mon 12 Tue 13 Wed 14 Thu 15 Fri 16 Sat
Tra
ffic
vo
lum
e [
Mb
ps]
Date (Mar. 2013, JST)
Home TXHome RX(a) home 2013
0
0.5
1
1.5
2
2.5
3
09 Sat 10 Sun 11 Mon 12 Tue 13 Wed 14 Thu 15 Fri 16 Sat
Tra
ffic
vo
lum
e [
Mb
ps]
Date (Mar. 2013, JST)
Pubilc TXPublic RX(b) public 2013
0
0.5
1
1.5
2
2.5
3
3.5
4
09 Sat 10 Sun 11 Mon 12 Tue 13 Wed 14 Thu 15 Fri 16 Sat
Tra
ffic
vo
lum
e [
Mb
ps]
Date (Mar. 2013, JST)
Office TXOffice RX(c) office 2013
0
5
10
15
20
25
30
35
40
45
28 Sat 01 Sun 02 Mon 03 Tue 04 Wed 05 Thu 06 Fri 07 Sat
Tra
ffic
vo
lum
e [
Mb
ps]
Date (Mar. 2015, JST)
Home TXHome RX(d) home 2015
0
0.5
1
1.5
2
2.5
3
28 Sat 01 Sun 02 Mon 03 Tue 04 Wed 05 Thu 06 Fri 07 Sat
Tra
ffic
vo
lum
e [
Mb
ps]
Date (Mar. 2015, JST)
Pubilc TXPublic RX(e) public 2015
0
0.5
1
1.5
2
2.5
3
3.5
4
28 Sat 01 Sun 02 Mon 03 Tue 04 Wed 05 Thu 06 Fri 07 Sat
Tra
ffic
vo
lum
e [
Mb
ps]
Date (Mar. 2015, JST)
Office TXOffice RX(f) office 2015
Figure 11: WiFi traffic volume: (a) home in 2013, (b) public in 2013, (c) office in 2013, (d) home in 2015, (e) public in 2015,and (f) office in 2015.
0
20
40
60
80
100
120
A H L | A H L | A H L
2013 2014 2015
1
2
34+
1
2
34+
1
2
34+
Perc
enta
ge
Figure 12: Percentage of the number of associated APs perday per user: All users (A), heavy hitters (H), and lightusers (L).
3.4.2 Access patterns of WiFi networks: locationWe first show the number of associated APs increases over
time, and that daily traffic volume per user does not corre-late to mobility pattern. Figure 12 illustrates the numberof associated APs per user per day for all users (A), heavyhitters (H), and light users (L). We find evidence that usersconnect to multiple WiFi APs in late years. In 2013, 70%of users only connect to one AP per day, but it decreases10% in three years. Moreover, we do not see a significantdifference between heavy hitters and light users. Thus, usertraffic volume does not correlate to the mobility patterns.
Next we detail that the main usage patterns of WiFi net-works are still only at home, and that the use of multipleAPs has been more common. Table 5 lists the breakdown ofthe number of associated ESSIDs per device in one day for allusers. The HPO column indicates the possible combinations
of WiFi network types; home (H), public (P), and other (O)networks. For example, HPO = 120 represents the numberof devices that connect to one home network, two public net-works, and no other network in one day. For users using oneAP, the percentage of using home WiFi decreases from 55%to 46%, though it is still dominant. About 10% of the usersdo not have any home WiFi network but connect to otherWiFi networks. They have no broadband wired links athome for WiFi offloading; however, they are aware of WiFitraffic offloading. For users using two APs, the top connec-tion pattern is one home and one other network. However,our manual inspection reveals that iOS devices connect moreto public networks than to other networks whereas Androiddevices connect oppositely. We also confirm low percentages(2.2-3.4%) of users who connect to more than three networksin one day. The maximum number of associated ESSIDs peruser per day is eight in our datasets: one home, four public,and three other networks.
We further show that the distribution of user connectiondurations to one WiFi network do not change over time.Figure 13 shows the complementary cumulative distributionof the consecutive time with the same AP in a log-log plot.All plots are characterized by long tails with a cutoff. Ninetypercent of the users connect for less than 12 hours for homenetworks, 8 hours for office networks, and 1 hour for publicnetworks. The results of home and office are consistent withtypical user activity by definition. No clear difference isobserved between the three datasets, meaning that the usagepatterns in connection duration do not change for any WiFinetwork type from 2013 to 2015.
3.4.3 Access patterns of WiFi networks: 2.4 and 5GHzRecent WiFi APs operate in two frequency bands: 2.4 and
5GHz. We find that deployment of 5GHz APs has been rapidfor public networks, but most home and office networks areat 2.4GHz. Deployment of 2.4GHz APs is wider than 5GHzbecause of their earlier availability on the market. However,
259
10-4
10-3
10-2
10-1
100
0.1 1 10 100
CC
DF
WiFi association time (hours)
public office home2013-15 2013-15 2013-15
Figure 13: CCDFs of WiFi connectionduration.
0
0.1
0.2
0.3
0.4
0.5
0.6
2013 2014 2015 2013 2014 2015 2013 2014 2015
Fra
ctio
n o
f a
sso
cia
ted
un
iqu
e 5
GH
z A
Ps home office public
Figure 14: Fractions of associatedunique 5GHz APs.
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
-90 -80 -70 -60 -50 -40 -30 -20 -10
PD
F
RSSI
home
public
Figure 15: PDFs of WiFi RSSI for asso-ciated APs (2015).
ESSIDsHPO 2013 2014 2015
per day
1100 54.7% 52.6% 46.4%010 3.0% 2.4% 2.4%001 10.5% 9.4% 9.2%
2
110 8.2% 10.0% 9.0%101 10.7% 12.9% 16.5%011 1.4% 1.3% 1.7%020 0.6% 0.3% 0.3%002 1.5% 1.8% 2.4%
3
102 1.8% 2.0% 2.7%120 1.9% 1.4% 1.4%111 2.2% 2.3% 3.4%012 0.3% 0.4% 0.6%021 0.4% 0.2% 0.3%003 0.3% 0.4% 0.5%
4+ — 2.3% 2.5% 3.2%
Table 5: Breakdown of number of associated APs (user·day):home (H), public (P), and other (O) networks.
the 2.4GHz spectrum is less robust against noise. We in-vestigate the deployment status of 5GHz APs at home, inoffices, and in public spaces. Figure 14 shows the fractionof unique associated 5GHz APs to the total unique asso-ciated APs for the three years. The fraction of 5GHz APsvaries across homes, offices, and public spaces. Still less than20% at home and office APs are 5GHz, i.e., most WiFi APsmainly operate in 2.4GHz. For public WiFi, we see thatmore than half are 5GHz APs in 2015. We believe that wideuse of 5GHz in public location is due to recent aggressiverollout of WiFi, while long device lifecycle means that homeand office have not yet needed to upgrade.
3.4.4 Quality of WiFi networks: signal strengthThere are several factors that affect the QoS/QoE of smart-
phones, in which the received signal strength indication (RSSI)is a common metric for quantifying the strength of the receiver-side signal in wireless communication. A larger RSSI repre-sents a stronger signal, and an RSSI larger than −70dBm isgenerally better for WiFi connectivity. For example, it hasbeen reported that the retransmission probability of a TCPsession is ≈ 10% for RSSI = −70dBm and rapidly increasesfor RSSI < −70dBm in a dataset measured at a conferenceWiFi network [41].
The signal strength at home is generally sufficient for net-work connectivity. We calculate the maximum RSSI of each
associated AP in 2.4GHz, and plot the probability densityof RSSIs of associated home and public networks in Fig-ure 15. Home networks are characterized by a bell shapewith a mean of −54dBm, consistent with those in Frenchhome networks [13, 14]. Three percent of the home networksreport a weaker RSSI than −70dBm. We also confirm thatthe shape of the plots of office networks is similar to that ofhome networks.
On the other hand, a small percentage of associated pub-lic APs indicate subpar quality. The plot of public WiFinetworks clearly shifts to smaller RSSIs around −60dBm.For narrowing the cell size in public WiFi, the mean signalstrength can be weaker than that of home WiFi networks.As the result, we confirm that 12% of associated publicWiFi networks indicate weaker signal strength (< −70dBm),likely causing poor QoS/QoE to the users. This low-qualityconnectivity is a reason to deploy more 5GHz APs in publicWiFi networks.
3.4.5 Quality of WiFi networks: channel usage
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0 2 4 6 8 10 12 14
PD
F
WiFi 2.4G Channel
(a) 2013
home APpublic AP
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0 2 4 6 8 10 12 14
PD
F
WiFi 2.4G Channel
(b) 2015
home APpublic AP
Figure 16: Associated 2.4GHz channels (2013 and 2015).
Finally, we investigate the efficacy of channel selectionof APs. In addition to the RSSI, channel selection of WiFinetworks is an important factor that affects the performanceof the WiFi networks. In an IEEE 802.11b/g/n network, 13channels are available in the 2.4GHz band, and two BSSIDsusing neighboring channels may interfere with each otherdue to the overlap of channel bandwidth (i.e., cross channelinterference). It is known that at least a five-channel intervalis necessary to avoid cross channel interference between two
260
channels. Some recent home-use APs have a function fordetecting and avoiding channel interference. In addition,public WiFi providers design and deploy their APs to avoidcross channel interference between their APs; however, theycan still interfere with other public APs.
We confirm that deployment of public APs is generallywell designed, and the concentration on the default channelhas been relaxed in home APs. Figure 16 shows the proba-bility density of the associated channels in the 2.4GHz bandfor Android devices. Devices connected mainly to Ch1, Ch6,and Ch11 in public WiFi APs, indicating that devices areset up to minimize interference. Such observation is alsorecently reported in [38, 6]. In 2013, the numerous connec-tions to Ch1 at home are, however, potentially causing morechannel interference. This poor setting is likely due to thelack of intelligent channel selection mechanisms at the APs.However, we observe improvement in avoiding possible in-terference in later years; WiFi channels at home are moredispersed and have less concentration on Ch1 in 2015.
3.5 Availability of public WiFi APsTo understand how much cellular traffic WiFi-available
users can offload to public WiFi networks, we next look atWiFi coverage, signal strength, and how much traffic theycan offload.
We first suggest a wide coverage of potentially availablepublic WiFi APs. We sum up all detected public networks(all) and detected public networks with a signal strengthstrong enough to be associated with (i.e., RSSI ≥ -70dBm)(strong), in a 5km cell. We confirm that 2.4GHz public WiFinetworks are widely deployed in the Greater Tokyo area inour data, though the coverage of our data is imperfect dueto the limited number of recruited users. In the downtownarea (Shinjuku, Shibuya area), for example, the number ofpublic WiFi networks with strong signal is over 10000 inone cell, while the number of detected networks is about45000 in one cell. Comparing the two datasets (2013 and2015), the number of cells with more than 100 strong 2.4GHzAPs increases from 102 to 139, and that with more than100 strong 5GHz APs is from 10 to 84. This observationindicates that deployment of 5GHz APs in public networkshas largely improved.
Next, we identify that WiFi-available users (§ 3.3.4) en-counter a few strong and available public WiFi APs, andthat the number of such APs increases. Figure 17 representsthe complementary cumulative distribution of the number ofdetected public APs per WiFi-available device per 10 min-utes in 2015. The plot labeled as “strong” represents thenumber of detected networks with strong maximum signalstrength, and“all” represents the number of all detected net-works. Both plots represent that a few users see many APsbut most see only a few. We observe that most users (90%)see fewer than 10 2.4GHz APs. As expected, the numberof strong WiFi networks is much smaller than all detectedWiFi networks. The data shows that 5GHz deployment isstill early, since only a few users (10%) could find strong5GHz APs, and only slightly more (30%) can find any 5GHz.However, this deployment is better than that in 2013; 10%for all APs and 3% for strong APs. Tail behavior of strong2.4GHz and strong 5GHz APs resembles each other, sug-gesting WiFi APs operating in both bands.
Finally, we infer that WiFi offloading to available publicWiFi networks can reduce 15-20% of daily cellular traffic for
Cell home Cell other WiFi home WiFi publictype % type % type % type %
2013
brows. 38.0 brows. 38.5 brows. 28.0 brows. 44.1social 7.3 comm. 7.7 social 6.8 social 4.0
comm. 6.2 social 7.6 comm. 4.3 life 3.3video 5.7 news 2.6 video 4.0 comm. 3.0news 2.0 video 2.1 news 3.5 news 2.9
2014
brows. 36.4 brows. 31.4 video 30.4 dload 22.5video 7.4 comm. 9.9 brows. 20.7 brows. 21.9
comm. 7.4 video 8.0 comm. 6.5 video 13.8social 6.3 news 6.6 news 6.0 life 4.9news 6.2 game 6.3 dload 4.7 health 3.2
2015
brows. 28.3 brows. 28.3 video 25.4 brows. 24.0video 11.0 comm. 12.7 brows. 20.0 video 19.6
comm. 9.5 video 12.0 dload 11.1 dload 9.9social 7.9 news 7.6 comm 7.4 life 4.1news 5.8 social 6.9 social 4.7 comm 3.6
Table 6: Top application categories, as ranked by RX trafficvolume.
WiFi-available users. As described in § 1, cellular providersdeploy free public WiFi service to their customers. Deviceswhose WiFi interfaces are enabled but not associated withany WiFi networks (i.e., WiFi-available user) can offloadtheir cellular traffic to the public WiFi networks of theirproviders. Examining of such available WiFi networks forWiFi-available users, we confirm that 60% of WiFi-availableusers have opportunities to connect to stable public WiFinetworks. We sum up cellular download traffic volume dur-ing these periods as possible offloading traffic, then concludethat 15-20% of daily cellular traffic volume for such WiFi-available users can be transferred to public WiFi networks.
3.6 Application breakdownWiFi offloading often requires application support, so we
now examine applications, their traffic volumes, and theirsupport for offload.
We first group popular applications into 26 categories inGoogle Play. Major application categories are: browser, so-cial networking (e.g., Facebook, Twitter), video and media(e.g., Youtube, Nicovideo), communication (e.g., Line, emailclients), news, gaming, music, travel, shopping, download-ing, entertainment (e.g., lottery, survey), tools (e.g., printer,speed test), productivity (e.g., online file storage service),and lifestyle (e.g., restaurant info, cooking). Note that thebrowser category includes web use; some web use may over-lap with other categories (such as videos and social networksthat are accessed from the browser instead of an applica-tion).
We distinguish application usage at home or other placeswith WiFi and cellular network interfaces. Table 6 and Ta-ble 7 show the percentage of the traffic volume (sum of RXand TX) generated from the five largest Android applicationcategories in each year, broken out by network type and lo-cation. We infer home locations of cellular networks withthe same classification technique described in § 3.4.1.
First of all, users are more likely to use high-bitrate appli-cations like video when they are on free networks like WiFi.The ratio of video traffic increases especially in WiFi net-works over the years, while we see that most common ap-plication categories in all scenarios are: browser, social net-working, video, and communications. Thus, watching videowith a smartphone has been a casual usage, though the usersunderstand that a rich and low-cost network environment is
261
10-6
10-5
10-4
10-3
10-2
10-1
100
1 10 100
CC
DF
Number of APs (public 2015)
2015
2.4GHz (all)2.4GHz (strong)
5GHz (all)5GHz (strong)
Figure 17: CCDFs of the number of de-tected public WiFi networks per deviceper 10 min.
0
0.2
0.4
0.6
0.8
1
09 11 13 15 17 19 21 23 25 0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
CD
F
PD
F
Date (Mar. 2015, JST)
(a)
(b)
CDF (all)
CDF (no home)
PDF (all)
Figure 18: Software update timing.
0
0.2
0.4
0.6
0.8
1
0.001 0.01 0.1 1 10 100
CD
F
Daily traffic / 3-days average
(a) 2014
potentiallycapped
others
0
0.2
0.4
0.6
0.8
1
0.001 0.01 0.1 1 10 100Daily traffic / 3-days average
(b) 2015
potentiallycapped
others
Figure 19: Effect of soft bandwidth cap.
Cell home Cell other WiFi home WiFi publictype % type % type % type %
2013
brows. 23.0 brows. 24.5 social 24.8 brows. 33.8video 16.1 social 15.1 brows. 18.4 social 9.1social 14.2 video 10.6 video 14.7 comm. 4.2
comm. 10.8 comm. 6.6 prod. 13.8 prod. 3.5prod. 2.1 prod. 3.0 comm 5.2 video 2.1
2014
brows. 31.0 brows. 28.5 prod. 39.5 brows. 41.8social 9.5 social 15.8 video 18.7 prod. 15.2prod. 9.0 comm. 11.5 brows. 15.7 life 6.9video 9.0 video 7.5 social 6.9 comm. 5.0
comm. 8.5 game 3.8 comm. 5.8 news 4.6
2015
brows. 33.7 brows. 29.7 brows 19.9 video 35.6comm 11.2 comm 14.5 prod 15.7 brows. 25.8social 11.0 social 11.3 video 12.7 prod 5.2game 9.9 life 5.9 social 10.5 comm. 5.0news 3.7 game 5.1 comm. 8.0 busi 4.6
Table 7: Top application categories, as ranked by TX trafficvolume.
preferable for this purpose. Furthermore, we observed an in-crease in video traffic in public WiFi networks. Thus, usersdemand for public WiFi networks is changing from simpleconnectivity-sensitive to more bandwidth-consuming applica-tions. Instead, most traffic of video and media category incellular at home is due to cellular-intensive users (withoutAPs at home).
We also confirmed some applications force users to useWiFi networks to reduce traffic volume in cellular networks.The percentage of the productivity category increases in up-load volume. In particular, in WiFi networks at home, thiscategory accounts for a high percentage of traffic volume dueto online storage software that uploads/downloads large filesonly if a WiFi interface is available. Thus, applications seemto play a major role in promoting WiFi offloading.
Finally, for light users, the contribution of video applica-tions to download traffic becomes small. We investigate theapplication usage of light users in 2015 (data omitted dueto space). The results are almost consistent with those forall users. A notable difference is a low contribution of videotraffic; it is not ranked in the top five categories in bothcellular and WiFi networks.
3.7 Software updateWhile § 3.6 looked at typical application use, in 2015 we
captured an unusual event where many iOS devices carriedout software upgrades. Apple only allows iOS upgrades on
WiFi, not cellular, by default. Thus, this event provides acase study of application-forced WiFi offloading.
The impact of the update on traffic volume is high; the sizeof the update is 565MB which is more than ten times largerthan the daily median download traffic volume (50.7MB),and it is more than half of the soft bandwidth cap limit.
Timing of software updates follows a typical flash crowd,with a large burst and a long tail. Figure 18 shows the time-of-day of updates in CDF and PDF since we first see an iOSupdate on March 10th, 2015 (JST). In two weeks, 58% of alliPhone devices are updated to iOS 8.2. Half users updatedin the first four days, with 10% of on the first day (cf. CDF(all)). We also confirm two broad peaks; one is the firstand second days (label (a) in the figure), and the other isthe first weekend (label (b)). Considering the percentage ofthe un-updated devices, the time for migrating to the newversion would have a longer tail.
We find that availability of WiFi is important to timely de-ploy updates; lack of WiFi can delay OS upgrades. In fact,only 14% of users without inferred home APs updated theirdevice OS. Moreover, the CDF of updates without homeAPs demonstrates that the updates of such users are de-layed. The median update delay of users with home WiFiand without is 3.5 days. For security-critical updates, thisdelay may leave users vulnerable longer.
Furthermore, we observed that some users without homeWiFi seem to go out-of-their way to access public WiFi toobtain the update. We inspect 19 updated devices that haveno inferred home APs. Eleven users without home AP up-date their devices via public WiFi, and two via office APs.This observation is unexpected but reasonable behavior thatreduces update costs.
3.8 Effect of soft bandwidth capJapanese cellular providers commonly limit heavy hitter’s
maximum bandwidth especially in peak hours. A typicalbandwidth cap begins after 1GB is received over the previousthree days. The download speed of users over the cap will belimited (e.g., 128kbps) during peak hours for the next fewdays. However, two cellular providers relaxed this policy inFebruary 2015, so we expect less impact from bandwidthcaps in 2015 compared to prior years.
We detect potentially capped users from their current andpast traffic volumes, and find that the ratio of such users issmall. The number of potentially capped users grows eachyear, from 0.5% in 2013, 0.8% in 2014, to 1.4% in 2015.
262
We can also see the effects of bandwidth limitation afterexceeding the cap. To observe this effect, we plot the ratiosof the daily traffic volume per user to the mean daily traf-fic volumes for the previous three days in 2014 and 2015 inFigure 19, where“potentially capped”represents users whosetraffic volume in the previous three days exceeds the thresh-old (= 1GB). The gap between potentially capped users andothers indicates the effect of the soft bandwidth cap. Thedownload traffic volume of capped users decreases due to thecap in 2014. Typically, 45% of capped users download lessthan half the mean of three-day traffic volume. On the otherhand, only 30% of other users download the same percent-age of traffic volume. Further manual inspection clarifiedthat 65% of capped users have no inferred home APs.
However, some users that exceed the threshold for band-width capping but appear to not be penalized with slower bitrates. A reason of this large traffic volume is likely downloadtraffic in off-peak hours.
By comparing the two figures, we confirm that the gap be-tween the capped users and others is small in 2015, showingthe change of the policy. The figure shows that the gap is0.15 in 2015 at the median while 0.29 in 2014. This changealso supports the reason of traffic increase of bandwidth-consuming applications in cellular networks from 2014 to2015 (Table 6).
4. IMPLICATIONSNext, we provide implications on the contribution of smart-
phone traffic on residential broadband traffic, connectivityand quality of WiFi networks, and bandwidth-consumingapplications.
4.1 Large impact of home WiFi offloadWe start by examining the impact of offloaded traffic vol-
ume on the total volume of residential broadband traffic.As shown in Figure 1, the nationwide cellular traffic volume(3G+LTE) corresponds to 20% of the estimated nationwideresidential wired broadband traffic volume [34], so a largeamount of smartphone traffic offloading will suggest an in-crease in broadband traffic. In our measurements, the me-dian cellular download traffic volume is 36MB/day and themedian WiFi download traffic is 51MB/day in 2015 (§ 3.2).Thus, 58% of smartphone traffic is WiFi; a 1.4:1 ratio ofWiFi-to-cellular traffic. This ratio is higher than that inCicso’s report (45% in 2014) [11]. As a result, we roughlyestimate 28% (= 20%*1.4) of the total residential broadbandtraffic volume as WiFi traffic by smartphones because 95%of WiFi traffic is at home (§ 3.4.1).
We also compare smartphone traffic volume with broad-band traffic volume per user. A different source of residentialbroadband traffic data reported that the median downloadtraffic volume of a residential customer per day in a JapaneseISP is 436MB/day as of 2015 [9]. We again infer with thisvolume that traffic share of one smartphone in a home net-work is 12% (= 51MB / 436MB).
Thus, the growth in offloaded WiFi traffic even for lightusers has had a large impact on legacy broadband ISPs, whereit is hidden in the traffic volume of heavy hitters.
4.2 Connectivity of WiFi networksTo understand the behavior of user that cannot be seen
in the network data, we had all our users fill out a ques-tionnaire. In the post questionnaire, we asked two questions
WiFi home office publicAP 13 14 15 13 14 15 13 14 15yes 70.4 72.9 78.2 31.6 25.6 28.0 44.9 47.9 53.6no 25.9 22.8 18.6 63.2 69.0 67.4 48.9 44.9 40.9
NA 3.7 4.3 3.2 5.1 5.4 4.6 6.1 7.2 5.5
Table 8: User survey: associated WiFi APs during the mea-surements (%: 2013/2014/2015).
about WiFi APs to users. The first question is “where didyou connect to WiFi APs in three locations (home, office,and public)?” (Table 8), to compare with the results of ourlocation-based analysis (§ 3.4.2). The second one is “whydid you not connect to WiFi APs in these three locations?”(Table 9), to understand possible causes of unavailability ofWiFi networks.
The answer to the first question shows that the percent-age of associated APs at home and in public increased whilethat for office network is low and stable. The percentages ofhome WiFi APs are consistent with our estimation (Table 5),but, high percentages of public WiFi differ from our esti-mation. This gap represents the difference between user’srecognition and actual connectivity; users think they havemore connectivity than they really do in public WiFi net-works. However, the results of the associated WiFi networksdemonstrate that the percentage of users with multiple APshas been increasing, and it is more than 40% in 2015 (§ 3.4).
From the measurement results and the answer to the sec-ond question (Table 9), we summarize the plausible reasonsfor the low availability of WiFi networks during daytimeas follows. (1) Users report that connecting private smart-phones to WiFi networks at offices (Bring your own device;BYOD) is still not common in Japan. It is supported byhigh percentages of the answer “there is no deployment ofAPs” in Table 9. In addition, a major reason of “Other” foroffice WiFi networks is due to the security policy that doesnot allow employees to connect their personal smartphoneto the office WiFi networks. Our traffic data showed a sig-nificant increase in office WiFi traffic volume (Figure 11)though the number of inferred office APs (Table 4) remainsstable during the three years.
(2) Configuring WiFi APs is difficult. Configuration prob-lems decrease in public WiFi networks over time in the sur-vey. Users report in our questionnaire that entering WiFipasswords into their mobile devices is often complicated.To reduce complexity, since 2013 WiFi APs from cellularproviders use SIM-based authentication [25] without userand password information. As a result, smartphone cus-tomers can use WiFi APs from their provider with no man-ual actions. Coupled with the deployment of more accesspoints (Figure 10), we believe this simplification in their usehelps increase the degree of WiFi offload.
(3) Many users have no incentive to use public WiFi net-works because the current quality of cellular networks isenough and are charged a flat fee per month. High percent-age of the answer “communication speed in LTE is enough”in Table 9 supports this claim. Nonetheless, we estimatedthat 15-20% of the traffic data can be offloaded if users ap-propriately configure the settings for public WiFi networks(§ 3.5).
(4) Finally, we found that security with public WiFi wasa significant concern, and that battery life was not.
263
Reason home office public13 14 15 13 14 15 13 14 15
No available APs 33 34 40 46 49 52 25 24 23Difficult to set up 32 27 21 16 15 11 31 31 25No configuration 48 35 32 33 25 22 43 31 29
Battery drain 18 14 15 16 9 7 25 18 13Failed 5 6 8 7 7 7 9 8 11
Security issue NA 6 14 NA 9 14 NA 15 35LTE is enough NA 25 21 NA 12 10 NA 22 23
Other 6 5 5 12 10 10 9 5 4
Table 9: User survey: reasons for unavailability of WiFi APs(%). Multiple answers were allowed.
4.3 Quality of public WiFi networksWe discuss how public WiFi networks can improve their
quality. The quality and accessibility of WiFi networks havebeen improved over the years. However, 12% of associatedpublic WiFi networks using 2.4GHz are characterized by alow RSSI (§ 3.4.4). Such low-quality network environmentsare likely a reason for users hesitating to use the public WiFinetwork, though over half of the associated public networkshave sufficient quality. To cope with this problem, deviceOSes should implement tests for the quality of a WiFi net-work before starting to offload to prevent the degradation ofuser QoE.
The coordinated deployment of public APs is a difficultproblem in reality because they use unlicensed bands. Tomitigate performance degradation due to channel interfer-ence among different public WiFi networks, deployment ofAPs that support multiple providers by announcing multi-ple ESSIDs from a single AP should be promoted, especiallyin downtown areas. In fact, we confirmed this type of APsin our dataset, by checking similar BSSIDs assigned to dif-ferent providers. It will be also cost-effective to use theseAPs for providing free WiFi APs to foreign visitors insteadof deploying new APs, towards the 2020 Olympic Games.Extended usage of such APs has already been discussed inthe context of providing free WiFi in disaster areas after theTohoku earthquake.
We also observed a significant deployment of 5GHz APs inpublic spaces (§ 3.4.3). 5GHz APs are more robust againstnoise and helpful in improving the quality of WiFi network[40, 46]. This migration also contributes to increases in avail-ability, use, and traffic in public WiFi networks. Further-more, the limited spatial and more reliable nature of 5GHzAPs may allow new types of applications to emerge, such asfine-grained targeting advertisement.
4.4 Bandwidth-consuming applicationsWe observed a large growth of bandwidth-consuming ap-
plication traffic (§ 3.6). Users with WiFi access downloadmore video (Table 7), suggesting that users understand thatrich-bandwidth and low-cost WiFi networks are available,and changing their usage accordingly. Similarly, other bandwidth-consuming applications contribute to WiFi offloading. Forexample, iOS shows a pop-up message to users to promotethe use of WiFi for software updates. We see a similarchange in upload use when WiFi is available. For example,WiFi has been largely used for the productivity category,which includes online file storage (Table 6).
This intelligent network connectivity management by soft-ware has two different perspectives. On one hand, it willbe more common and necessary to cope with bandwidth-
consuming applications in the future, particularly to avoidbandwidth caps. On the other hand, it causes another adap-tive behavior of users. Some users rely on public WiFi net-works for their software updates (§ 3.7). Furthermore, weobserved an increase in video traffic even in public WiFinetworks. User demand is towards casual use of bandwidth-consumption applications in public WiFi networks. Thistrend may introduce a change of the economic cost and theQoS model (i.e., traffic engineering) of public and free WiFinetworks, because low-cost connectivity is the main advan-tage with current WiFi networks.
5. RELATED WORKCellular smartphone usage: Quantifying the diverse
usage of smartphones has been one of the hot topics in traf-fic measurement and analysis. Some studies have pointedout that the flow-level performance of the 3G/LTE cellularnetwork is affected by many aspects such as the differencesin the cellular providers, types of devices, transport pro-tocols, and mobility [32, 28, 29, 27, 37, 19, 42, 17]. Theapplication usage and resulting performance in cellular net-works also indicate large diversity [47, 16, 28, 51, 44, 27].In particular, the usage pattern of applications in smart-phones depends on the mobility and geographical region ofusers [47, 51, 52]. Recent studies addressed the impact ofuser measurement environment on network performance [21,43]. There are several approaches to characterize this di-versity such as application identification [50, 36], app storeusage [49], and geolocation mapping [3, 48]. Another di-rection of current research is to quantify and model energyconsumption of smartphones [27, 15, 7]. Regarding band-width caps, a recent experiment concerning cellular servicepricing shows that time-dependent pricing helps reduce peaktraffic volume in cellular networks [30].
WiFi smartphone usage: Some studies focused on WiFismartphone usage in campus networks [26, 22, 8], city-levelnetworks [1], or networks in public transportations [24]. Theyargued that typical applications of such WiFi networks wereHTTP, more specifically video, social networking, and soft-ware updates. In addition, the deployment of WiFi APs andtheir interference are discussed [2].
Throughput/delay performance of WiFi and 3G/LTEnetworks: Several performance metrics have been com-pared using vehicle-based [4, 20, 12] and speedtest-based [45]measurements. They showed that the throughput and de-lay performance of WiFi networks are basically better thanthose of cellular networks, and they are also highly vari-able depending on the environment, different from broad-band wired access. Furthermore, in the context of trafficoffloading, some studies evaluated the performance gain ofWiFi offloading on the basis of WiFi APs discovered usingvehicle based measurements [31, 23].
WiFi networks: Home WiFi usage has been analyzedin detail [33, 39, 13, 14]. An earlier report pointed out a lowpercentage of traffic volume from mobile devices at homein 2009 [33]. Also, the performance [40, 46], the traffic ra-tio [33], and the availability [13, 14] of home WiFi networkshave been intensively studied. A large analysis of ESSIDsfocused on the social relationships of users [5]. A recentlarge-scale survey of WiFi networks revealed application us-age and channel interference in the wild [6].
The main difference between our work and others is to pro-vide a comprehensive view of smartphone usage in cellular
264
and WiFi networks from measurements with over 1500 users,for better understanding the adoption of smartphone usersto avaialble WiFi networks. We previously explained thepreliminary results of WiFi traffic offloading performancebased on a two-day measurement with 450 Android users [18].However, the measurement period was short and the avail-able information was limited mainly to traffic volume with-out application and geographical information. For the cur-rent study, we analyzed three 15-day-long detailed datasetsand provided more comprehensive results including the dif-ferences regarding device OS, location, and applications.
6. CONCLUDING REMARKSWe measured WiFi and 3G/LTE smartphone usage from
more than 1500 Android/iOS smartphone users in the GreaterTokyo area for two weeks each in 2013, 2014, and 2015. Ourresults showed that smartphone users select WiFi and cel-lular networks based on several tradeoffs. In particular, wehighlighted slow but clear adoption of available WiFi envi-ronment at home, in offices, and in public spaces. Lightusers have been offloading their traffic to WiFi networks,while heavy hitters totally rely on WiFi networks for theirtraffic offload. Moreover, we confirmed that wide deploy-ment of WiFi APs enables users to obtain more opportu-nities to offload their traffic to WiFi networks. Users whohave access to public WiFi send more traffic, and WiFi-availability results in heavier use of video and is required forlarge software updates. This adoption indicates that usersrecognize public WiFi networks as a part of the commonnetwork infrastructure.
AcknowledgmentsWe thank our shepherd Dina Papagiannaki, the IMC re-viewers, Kenjiro Cho, Romain Fontugne, John Heidemann,Johan Mazel, and Joao Taveira Araujo for their valuablecomments, and the Ministry of Internal Affairs and Com-munications, Japan for its support.
7. REFERENCES[1] M. Afanasyev, T. Chen, G. M. Voelker, and A. C.
Snoeren. Analysis of a mixed-use urban WiFi network:When metropolitan becomes neapolitan. In IMC’08,pages 85–98, Vouliagmeni, Greece, Oct 2008.
[2] A. Akella, G. Judd, S. Seshan, and P. Steenkiste.Self-management in chaotic wireless deployments. InMobiCom’05, pages 185–199, Cologne, Germany, Aug2005.
[3] M. Balakrishnan, I. Mohomed, andV. Ramasubramanian. Where’s that phone?:Geolocating IP addresses on 3G networks. In IMC’09,pages 294–300, Chicago, IL, Nov 2009.
[4] A. Balasubramanian, R. Mahajan, andA. Venkataramani. Augmenting mobile 3G usingWiFi. In MobiSys’10, pages 209–222, San Francisco,CA, Jun 2010.
[5] M. V. Barbera, A. Epasto, A. Mei, V. C. Perta, andJ. Stefa. Signals from the crowd: Uncovering socialrelationships through smartphone probes. In IMC’13,pages 265–276, Barcelona, Spain, Oct 2013.
[6] S. Biswas, J. Bicket, E. Wong, R. Musaloiu-E,A. Bhartia, and D. Aguayo. Large-scale measurements
of wireless network behavior. In SIGCOMM’15, pages153–165, London, UK, Aug 2015.
[7] X. Chen, N. Ding, A. Jindal, Y. Hu, M. Gupta, andR. Vannithamby. Smartphone energy drain in thewild: Analysis and implications. In SIGMETRICS’15,pages 151–164, Portland, OR, Jun 2015.
[8] X. Chen, R. Jin, K. Suh, B. Wang, and W. Wei.Network performance of smart mobile handhelds in auniversity campus WiFi network. In IMC’12, pages315–328, Boston, MA, 2013.
[9] K. Cho. Broadband traffic report. In InternetInfrastructure Review, volume 28, pages 28–33.Internet Initiative Japan, Aug 2015.
[10] K. Cho, K. Fukuda, H. Esaki, and A. Kato. Observingslow crustal movement in residential user traffic. InCoNEXT’08, page p.12, Madrid, Spain, Dec 2008.
[11] Cisco. Cisco visual networking index. http://www.cisco.com/c/en/us/solutions/collateral/
service-provider/visual-networking-index-vni/
white_paper_c11-520862.html.
[12] P. Deshpande, X. Hou, and S. R. Das. Performancecomparison of 3G and metro-scale WiFi for vehicularnetwork access. In IMC’10, pages 301–307, Melbourne,Australia, Nov 2010.
[13] L. DiCioccio, R. Teixeira, and C. Rosenberg.Characterizing home networks with homenet profiler.Technical report, Technicolor Technical ReportCR-PRL-2011-09-0001, 2011.
[14] L. DiCioccio, R. Teixeira, and C. Rosenberg.Measuring home networks with homenet profiler. InPAM’13, pages 176–186, Hong Kong, China, Mar2013.
[15] N. Ding, D. Wagner, X. Chen, A. Pathak, Y. Hu, andA. Rice. Characterizing and modeling the impact ofwireless signal strength on smartphone battery drain.In SIGMETRICS’13, pages 29–40, Pittsburgh, PA,Jun 2013.
[16] H. Falaki, R. Mahajan, S. Kandula,D. Lymberopoulos, R. Govindan, and D. Estrin.Diversity in smartphone usage. In MobiSys’10, pages179–194, San Francisco, CA, Jun 2010.
[17] P. Fiadino, M. Schiavone, and P. Casas. Vivisectingwhatsapp in cellular netoworks: Servers, flows, andquality of experience. In TMA’15, pages 49–63,Barcelona, Spain, Apr 2015.
[18] K. Fukuda and K. Nagami. A measurement of mobiletraffic offloading. In PAM’13, pages 73–82, HongKong, China, Mar 2013.
[19] Z. Gao, A. Venkataramani, J. F. Kurose, andS. Heimlicher. Towards a quantitative comparison oflocation-independent network architectures. InSIGCOMM’14, pages 259–270, Chicago, IL, Aug 2014.
[20] R. Gass and C. Diot. An experimental performancecomparison of 3G and Wi-Fi. In PAM’10, pages71–80, Zurich, Switzerland, Apr 2010.
[21] A. Gember, A. Akella, J. Pang, A. Varshavsky, andR. Caceres. Obtaining in-context measurements ofcellular network performance. In IMC’12, pages287–317, Boston, MA, Nov 2013.
265
[22] A. Gember, A. Anand, and A. Akella. A comparativestudy of handheld and non-handheld traffic in campusWi-Fi networks. In PAM’11, pages 173–183, Mar 2011.
[23] B. Han, P. Hui, V. A. Kumar, M. V. Marathe, G. Pei,and A. Srinivasan. Cellular traffic offloading throughopportunistic communications: A case study. InCHANTS’10, pages 31–38, Chicago IL, Sep 2010.
[24] J. Hare, L. Hartung, and S. Banerjee. Beyonddeployments and testbeds: Experiences with publicusage on vehicular WiFi hotspots. In MobiSys’12,pages 393–405, Low Wood Bay, UK, Jun 2012.
[25] H. Haverinen and J. Salowey. EAP-SIMAuthentication. RFC4186, Jan 2006.
[26] T. Henderson, D. Kotz, and I. Abyzov. The changingusage of a mature campus-wide wireless network. InMobiCom’04, pages 187–201, Philadelphia, PA, 2004.
[27] J. Huang, F. Qian, Y. Guo, Y. Zhou, Q. Xu, Z. M.Mao, S. Sen, and O. Spatscheck. An In-depth study ofLTE: Effect of network protocol and applicationbehavior on performance. In SIGCOMM’13, pages363–374, Hong Kong, China, Aug 2013.
[28] J. Huang, Q. Xu, B. Tiwana, Z. M. Mao, M. Zhang,and P. Bahl. Anatomizing application performancedifferences on smartphones. In MobiSys’10, pages165–178, San Francisco, CA, Jun 2010.
[29] K. Jang, M. Han, S. Cho, H.-K. Ryu, J. Lee, Y. Lee,and S. Moon. 3G and 3.5G wireless networkperformance measured from moving cars andhigh-speed trains. In MICNET’09, pages 19–24,Beijing, China, Oct 2009.
[30] C. Joe-Wong, S. Ha, S. Sen, and M. Chiang. Domobile data plans affect usage? results from a pricingtrial with isp customers. In PAM’15, pages 96–108,New York, NY, Mar 2015.
[31] K. Lee, I. Rhee, J. Lee, S. Chong, and Y. Yi. Mobiledata offloading: How much can WiFi deliver? InCoNEXT’10, page 12, Philadelphia, PA, Dec 2010.
[32] X. Liu, A. Sridharan, S. Machiraju, M. Seshadri, andH. Zang. Experiences in a 3G network: interplaybetween the wireless channel and applications. InMobiCom’08, pages 211–222, San Francisco, CA, Sep2008.
[33] G. Maier, F. Schneider, and A. Feldmann. A first lookat mobile hand-held device traffic. In PAM’10, pages161–170, Zurich, Switzerland, Apr 2010.
[34] Ministry of Internal Affairs and Communications.Information and communications statistics database.http://www.soumu.go.jp/johotsusintokei/field/
tsuushin06.html.
[35] Ministry of Internal Affiars and Communications.SAQ2 Japan Project: Action Plan for EstablishingICT Use Environment for Foreign Tourists. http://www.soumu.go.jp/main_content/000306547.pdf,2014.
[36] S. Miskovic, G. M. Lee, Y. Liao, and M. Baldi.AppPrint: automatic fingerprinting of mobileapplications in network traffic. In PAM’15, pages57–69, New York, NY, Mar 2015.
[37] A. Nikravesh, D. R. Choffnes, E. Katz-Bassett, Z. M.Mao, and M. Welsh. Mobile network performancefrom user devices: A longitudinal, multidimensionalanalysis. In PAM’14, pages 12–22, Mar 2014.
[38] Ofcom. Study on the use of Wi-Fi for metropolitanarea applications. http://stakeholders.ofcom.org.uk/market-data-research/other/
technology-research/2013/wifi-met-area/, 2013.
[39] K. Papagiannaki, M. Yarvis, and W. S. Conner.Experimental characterization of home wirelessnetworks and design implications. In INFOCOM’06,page 13, Barcelona, Spain, Apr 2006.
[40] A. Patro, S. Govindan, and S. Banerjee. Observinghome wireless experience through wifi aps. InMobiCom’13, pages 339–350, Miami, FL, Sep 2013.
[41] M. Rodrig, C. Reis, R. Mahajan, D. Wetherall, andJ. Zahorjan. Measurement-based characterization of802.11 in a hotspot setting. In E-WIND’05, pages5–10, Philadelphia, PA, Aug 2005.
[42] P. Romirer-Maierhofer, M. Schiavone, andA. D’Alconzo. Device-specific traffic characterizationfor root cause analysis in cellular networks. InTMA’15, pages 49–63, Barcelona, Spain, Apr 2015.
[43] M. Z. Shafiq, L. Ji, A. X. Liu, J. Pang,S. Venkataraman, and J. Wang. A first look at cellularnetwork performance during crowded events. InSIGMETRICS’13, pages 17–28, Pittsuburgh, PA, Jun2013.
[44] M. Z. Shafiq, L. Ji, A. X. Liu, and J. Wang.Characterizing and modeling internet traffic dynamicsof cellular devices. In SIGMETRICS’11, pages305–316, San Jose, CA, Jun 2011.
[45] J. Sommers and P. Barford. Cell vs. WiFi: On theperformance of metro area mobile connections. InIMC’12, pages 301–314, Boston, MA, Nov 2012.
[46] S. Sundaresan, N. Feamster, and R. Teixeira.Measuring the performance of user traffic in homewireless networks. In PAM’15, pages 305–317, NewYork, NY, Mar 2015.
[47] I. Trestian, S. Ranjan, A. Kuzmanovic, and A. Nucci.Measuring serendipity: Connecting people, locationsand interests in a mobile 3G network. In IMC’09,pages 267–279, Chicago, IL, Nov 2009.
[48] S. Triukose, S. Ardon, A. Mahanti, and A. Seth.Geolocating IP addresses in cellular data networks. InPAM’12, pages 158–167, Vienna, Austria, 2012.
[49] N. Viennot, E. Garcia, and J. Nieh. A measurementstudy of google play. In SIGMETRICS’14, pages221–233, Austin, TX, Jun 2014.
[50] X. Wei, L. Gomez, L. Neamtiu, and M. Faloutsos.Profiledroid: Multi-layer profiling of androidapplications. In MobciCom’12, page 12, Istanbul,Turkey, Aug 2012.
[51] Q. Xu, J. Erman, A. Gerber, Z. Mao, J. Pang, andS. Venkataraman. Identifying diverse usage behaviorsof smartphone apps. In IMC’11, pages 329–344,Berlin, Germany, Nov 2011.
[52] Z. Zhu, G. Cao, R. Keralapura, and A. Nucci.Characterizing data services in a 3G network: Usage,mobility and access issues. In ICC’11, page 6, Kyoto,2011.
266
